Music Summary using Key Phrases

Abstract

As the magnitude and use of multimedia databases grows rapidly, efficient ways to automatically find the “gist” of the contents becomes a necessity. This work addresses the problem of summarizing music, specifically songs of rock or pop genre. We assert that a typical song can be summarized by one or more representative “key phrases”. Our goal then is to identify the reoccurring temporal patterns in the audio signal. We investigate two approaches to summarization. In the first, we divide a song into fixed-length segments and cluster these based on a similarity measure. The key phrase is extracted by choosing the most frequent cluster. In the second approach, Hidden Markov Models (HMMs) are used to discover the structure of the music. Both approaches use the Mel-frequency cepstrum as the feature. The two proposed approaches were evaluated on a set of Beatles songs. The clustering method consistently achieved good performance, while the HMM method produced mixed results.